Deepseek Releases V4-Flash-Vision-Exp Vision Model

Chinese artificial intelligence company Deepseek has launched V4-Flash-Vision-Exp, an experimental multimodal model that integrates image processing with its existing text capabilities. Built on top of the Deepseek-V4-Flash base model, this new release retains the original model's reasoning and world knowledge while introducing visual understanding. On Deepseek's internal multimodal agent benchmarks, the experimental vision model nearly matches or sometimes exceeds the performance of Opus 4.8, making it a highly competitive option for complex agentic tasks.

For developers building visual agent workflows, the model offers broad compatibility and flexible integration options. It works out of the box with OpenAI's Chat Completions and Responses APIs, as well as Anthropic's Messages endpoint. Deepseek has also updated its Harness framework to version 0.1.1 to support the new model. Practitioners can send images to the model in three ways: via Base64 encoding, through public URLs up to 32 MiB, or by using Deepseek's new, free Files API, which supports uploads up to 64 MiB and allows files to be referenced by ID across multiple requests.

The model processes JPEG, PNG, GIF, and WebP formats, identifying the file type directly from the content rather than the extension. It can handle up to 600 images in a single request, with a maximum edge length of 8,192 pixels per side, which drops to 4,096 pixels when a request contains 15 or more images. Images are normalized to roughly 800 x 800 pixels during processing, costing a maximum of 384 tokens per image under standard V4-Flash pricing. Developers can also use an optional detail field to downscale images to 512 x 512 pixels, further reducing token consumption for tasks that do not require fine visual resolution.